SSD detection

deeplearning

---

anchor box生成

以38*38的feature map为例
step 1
初始化x,y

y=array([[0,0.....0],
         [1,1.....1],
         .....
         [38,38...38]])
x=array([[0,1,....38],
         [0,1,....38],
         ...
         [0,1,....38]])

将x,y归一到0-1

y=(y+offset)*step/img_shape[0] #step为img_shape[0]/feature_shape[0]

初始化w,h

ground truth预处理

计算ymin,ymax,xmin,xmax
计算IoU
保留大于threshold大小的anchor box作为正样本

计算loss

SSD目标函数分为两个部分：对应默认框的位置loss（loc）和类别置信度loss（conf）。定义

$$x_{ij}^{p}=\left\{ 1,0 \right\}$$ 为第i个默认框和对应的第j个ground truth box，相应的类别为p。目标函数定义为：
http://www.zhihu.com/equation?tex=L%28x%2Cc%2Cl%2Cg%29%3D%5Cfrac%7B1%7D%7BN%7D%28L_%7Bconf%7D%28x%2Cc%29%2B%5Calpha+L_%7Bloc%7D%28x%2Cl%2Cg%29+%29